基于广义中心差分卷积和空间分布机制的CG图像检测

CG Image Detection Based on Generalized Center Difference Convolution and Spatial Layout Mechanism

针对数字取证和司法鉴定领域中计算机生成图像检测技术日益增长的现实需求,提出一种基于广义中心差分卷积和空间分布机制的计算机生成图像检测网络。首先,设计了一个包含三个并行独立分支的相关性特征提取模块;随后,将三个分支的输出经串接后输入到通道注意力机制子模块;最后,使用5个附带空间分布机制的深度卷积模块进一步学习图像的分层表示来进行最终决策。在SPL2018和DSToK两个公共数据集上的检测准确率可达94.76%和95.38%,相比最好的对比方法对生成图像的检测准确率提高了3.12%和3.23%。消融实验验证了网络中各模块对于模型整体检测效果的贡献。最后,验证了该网络对JPEG压缩和加性噪声的鲁棒性,即使对质量因子为60压缩后的图像,检测准确率仍可达84%以上。提高了模型的检测准确率及鲁棒性。

A computer-generated image(CGI)detection network based on generalized center difference convolution(GCDC)and spatial layout mechanism(SLM)is proposed to meet the growing demand of computer-generated image detection technology in the fields of digital forensics and judicial expertise.Firstly,a correlation feature extraction module containing three parallel independent branches is designed.Next,the outputs of the three branches are concatenated and then fed into a channel attention mechanism submodule.Finally,five successive convolutional modules with spatial layout mechanism modules are utilized to learn higher level hierarchical representation for making decision.The detection accuracy on the SPL2018 and DSToK public datasets can reach 94.76%and 95.38%,which is 3.12%and 3.23%higher than the best comparison method in detecting generated images.The detection accuracy on the SPL2018 and DSToK public datasets can reach 94.76%and 95.38%,which is 3.12%and 3.23%higher than the best comparison method in detecting generated images.The ablation experiment verifies the contribution of each module in the network to the overall detection performance of the model.Finally,the robustness of the network to JPEG compression and additive noise is verified,and even for compressed images with a quality factor of 60,the detection accuracy can still reach over 84%.